Cross-Domain Radar Gesture Generation via Diffusion Models With Transfer Guidance and Physical Priors

Le résumé fourni par la source

Radar-based gesture recognition has shown great potential in human-computer interaction due to its strong privacy-preserving capabilities. However, practical deployment is severely constrained by factors such as device installation limitations, environmental variability, and high data acquisition costs. These challenges lead to sample scarcity and cross-domain distribution shifts, which significantly degrade both generation and recognition performance. Existing generative methods suffer from mode collapse under extreme data scarcity and produce samples lacking kinematic physical constraints, compromising structural plausibility. To address these challenges, this paper proposes Transfer-Guided Diffusion Process with Physical Prior (TGDP-P), a diffusion-based radar gesture generation method that synergistically integrates transfer guidance with physical priors. During the reverse denoising process of the diffusion model, TGDP-P incorporates a transfer-guidance mechanism and physical prior constraints. Specifically, a density-ratio correction term is introduced to efficiently bridge the distribution gap between source and target domains, while time-frequency ridge regularization enforces kinematic consistency in the generated samples. Ablation studies confirm that the physical prior is critical to generation quality, and the synergy between distribution guidance and physical constraints enables TGDP-P to generate samples with both high fidelity and strong physical plausibility. Experiments on a simulated dataset and the public UWB-Gestures benchmark demonstrate that the proposed method maintains superior performance under extreme data scarcity. With 100 samples from the target domain, TGDP-P achieves an F1-score of 0.7970 and improves classification accuracy by 2.60 percentage points compared to the baseline. Furthermore, quantitative evaluation across four representative cross-domain scenarios validates the strong generalization capability of TGDP-P, offering a high-fidelity and robust data generation solution for radar-based gesture recognition under resource-constrained conditions.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.